★  Best Track Paper Award — SEDSI 2025  ★

Research · SEDSI 2025 Conference Paper

The First Five Seconds

What makes people trust AI before they've had time to think. A separate study from the same dataset revealed something the dissertation didn't set out to ask: the type of legitimacy that builds trust shifts entirely depending on how much is riding on the decision. The same AI. Different stakes. Completely different logic.

Dr. Albert “Trey” H. Harper III, with Dr. Farnaz Ghashami & Dr. David Gefen · LeBow College of Business, Drexel University · SEDSI 2025

The dissertation I defended in February 2025 asked a broad question: what drives someone to advocate for generative AI? The paper that came out of the same research project — this one, awarded Best Track Paper at SEDSI 2025 — asked something sharper.

It asked: what happens in the first moment?

Not the considered judgment. Not the accumulated experience. The snap reaction — the trust or suspicion that forms before someone has had enough time to deliberate. Researchers call this swift trust and swift distrust. And it turns out the rules governing swift trust are fundamentally different from the rules governing trust that builds over time.

More importantly for anyone deploying AI today: those rules change depending on what's at stake.

Two studies, one dataset, a different question

The dissertation examined trust and distrust as mature constructs — shaped by experience, knowledge, accumulated use. It asked how legitimacy, risk, and advocacy intentions interact across the full arc of a user relationship with GenAI.

This paper asked a more constrained version of that question. Imagine you've just been handed a new AI tool and asked to use it now, today, for a real task. Before you've built any track record with it. Before you've seen it fail or succeed more than once. What makes you trust it in that moment — or not?

Same 715 participants, same low-risk versus high-risk scenario design (document summarization vs. financial advice), same legitimacy framework. Different construct at the center: swift trust — the kind of trust that forms in temporary, uncertain conditions, based on limited information and generalized expectations.

What is swift trust, exactly?

The concept comes from organizational research on temporary teams — groups of people thrown together to accomplish a time-sensitive task who don't have the luxury of building relationships slowly. They have to decide, fast, whether to rely on each other.

GenAI creates exactly that situation at the individual level. Someone is assigned a task. They open a tool they've never used — or haven't used much. They have to decide, right now, whether to rely on what it outputs. There's no trial period. No accumulated evidence. Just the tool, the task, and a judgment call.

Swift distrust is the flip side: the immediate suspicion that forms under the same conditions, rooted in concern about potential harm rather than confidence in capability.

The central finding: the stakes change the rules

This is the result the paper was built around, and it's worth stating plainly before walking through the details.

When the stakes are low — someone using AI to summarize a document — swift trust forms primarily through pragmatic legitimacy: does this thing actually work? Does it solve my problem? Does it deliver visible, immediate value?

When the stakes are high — someone using AI to generate financial advice — the pragmatic question almost disappears. Instead, swift trust depends on cognitive legitimacy (is this credible? is it aligned with how respected players in my industry operate?) and regulative legitimacy (is this operating within a recognized legal and regulatory framework?).

The shift isn't gradual. It's discrete. Low stakes: show me it works. High stakes: show me it's sanctioned.

The distrust finding everyone misses

Most AI rollout strategies are built around building trust. The procurement team emphasizes benefits. The change management team runs trainings. Leadership sends the “we're committed to responsible AI” memo.

This study points to a gap in that playbook.

Swift distrust — the immediate suspicion that forms at first encounter — was consistently driven by a single factor across both risk scenarios: regulative legitimacy. Or more precisely, its absence.

Regulatory compliance, in other words, doesn't primarily build trust. It primarily prevents distrust. Those are different jobs — and conflating them leads to under-investment in governance communication at exactly the moment it matters most: the first encounter.

The practical implication is sharper than it might appear. You can have an AI system that is pragmatically excellent, cognitively credible, and ethically well-designed — and still generate immediate user suspicion if the governance and compliance story is invisible at the point of first use.

Why normative legitimacy didn't show up

One finding that surprised me: normative legitimacy — whether the AI aligns with ethical standards and the common good — wasn't a significant driver of swift trust or swift distrust in either scenario.

This isn't a statement that ethics don't matter. They do. But ethics are a long-game legitimacy signal. Evaluating whether a technology genuinely reflects good values requires time, reflection, and accumulated exposure. It's a judgment that forms over months, not minutes.

Swift trust is a first-minutes phenomenon. In the first minutes, people are asking more immediate questions: does this work (pragmatic), is this credible (cognitive), is this sanctioned (regulative)? The ethical question — is this right? — matters, but it comes later.

For organizations that have invested heavily in ethical AI frameworks and governance commitments, this is worth sitting with. Your ethics story may be your most important long-term legitimacy asset — but it probably isn't your first-encounter trust signal. Lead with something else at rollout. Let the ethics story compound over time.

The credibility constant

What unified both scenarios — what didn't change when the stakes changed — was the role of cognitive legitimacy in driving advocacy. In both the low-risk and the high-risk contexts, users who perceived the AI as credible, associated with respected industry leaders, and aligned with established practice were more likely to recommend it to others.

This connects to the dissertation's broader finding about advocacy: people champion AI they believe in, and belief tracks credibility more than it tracks any other single legitimacy signal.

What this means for AI rollout design

Taken together, the findings suggest that the opening moment of an AI deployment should be designed differently depending on what the AI is being asked to do.

For low-stakes applications — productivity tools, content generation assistants, summarization — lead with utility. Make the value visible immediately. Let people experience pragmatic legitimacy firsthand before you make any other claim about the system. Governance and compliance matter here too, but they're supporting signals rather than primary ones.

For high-stakes applications — financial decision support, medical information, legal analysis, anything with significant downstream consequences — the playbook reverses. Utility is secondary at first encounter. What matters is whether users can see credible institutional alignment and clear regulatory grounding. If those signals are absent when someone opens the tool for the first time, distrust forms before you get a chance to demonstrate value.

In practice: the AI system's compliance posture and industry credentialing need to be visible at the interface level in high-stakes contexts — not buried in a vendor whitepaper or an IT governance document no user will ever read.

The limitation worth naming

Swift trust is a first-encounter phenomenon, and this study measures exactly that — a moment. What happens to trust over time, as users accumulate real experience with a system, is a different question this study wasn't designed to answer.

It's also worth acknowledging that the high-risk and low-risk categories (financial advice and document summarization) are reasonable proxies but not universal. Different users, cultures, and industries will place different thresholds on what counts as “high stakes.” A financial analyst has a different relationship to financial advice AI than a generalist professional. The legitimacy signals that matter may be calibrated differently within specific domains.

Future work should track legitimacy perceptions longitudinally and test the framework across industry-specific contexts where the risk calculus is more precisely defined.

Where this sits in the larger argument

The dissertation's finding was that trust and distrust are separate systems — that building one doesn't automatically dismantle the other, and that they require targeted, independent interventions.

This paper adds a temporal layer to that argument. Even within the trust system alone, the signals that matter shift depending on when in the user journey you're measuring and what context you're measuring them in.

First-encounter trust — swift trust — doesn't just respond to the same legitimacy signals as considered trust, faster. It responds to different signals entirely, depending on what's at stake. That matters for deployment strategy. It matters for how you sequence the legitimacy case you make to users. And it matters for how organizations think about governance visibility at the moment it counts most: before users have had time to form any opinion at all.

The first five seconds aren't neutral. Design them accordingly.